Comparison Gallery

Reconstruction

We provided the reconstruction results of baseline methods[Bagher et al., Sun et al., Hu et al.] and our networks(NPs model with different setups, hypernet). We rendered the reconstruction BRDF under three different environment maps(St, Grace, Uffizi).

Links	Description
BagHer-Grace BagHer-Uffizi BagHer-St	[Bagher et al.]
Hu-Grace Hu-Uffizi Hu-St	[Hu et al.]
Sun-Grace Sun-Uffizi Sun-St	[Sun et al.] with 1 diffuse lobe and 5 specular lobes
Hypernet-Grace Hypernet-Uffizi Hypernet-St	our hypernet with 4 times log mappings, 7-dimensional latent space, and mean aggregator
FinalNPs-Grace FinalNPs-Uffizi FinalNPs-St	our NPs model with 4 times log mappings, 7-dimensional latent space, and mean aggregator
NPsLOG2DIM7-Grace NPsLOG2DIM7-Uffizi NPsLOG2DIM7-St	our NPs model with 2 times log mappings, 7-dimensional latent space, and mean aggregator
NPsLOG3DIM7-Grace NPsLOG3DIM7-Uffizi NPsLOG3DIM7-St	our NPs model with 3 times log mappings, 7-dimensional latent space, and mean aggregator
NPsLOG4DIM2-Grace NPsLOG4DIM2-Uffizi NPsLOG4DIM2-St	our NPs model with 4 times log mappings, 2-dimensional latent space, and mean aggregator
NPsLOG4DIM3-Grace NPsLOG4DIM3-Uffizi NPsLOG4DIM3-St	our NPs model with 4 times log mappings, 3-dimensional latent space, and mean aggregator
NPsLOG4DIM4-Grace NPsLOG4DIM4-Uffizi NPsLOG4DIM4-St	our NPs model with 4 times log mappings, 4-dimensional latent space, and mean aggregator
NPsLOG4DIM5-Grace NPsLOG4DIM5-Uffizi NPsLOG4DIM5-St	our NPs model with 4 times log mappings, 5-dimensional latent space, and mean aggregator
NPsLOG4DIM6-Grace NPsLOG4DIM6-Uffizi NPsLOG4DIM6-St	our NPs model with 4 times log mappings, 6-dimensional latent space, and mean aggregator

Importance Sampling

We compared three importance sampling strategies(ours, cosine-weighted, GGX-based) under four environment maps.

ImportanceSampling-Grace
ImportanceSampling-Uffizi
ImportanceSampling-St
ImportanceSampling-envmap

Moreover, to validate that our NICE network works well on interpolated materials, we interpolated each materials in the MERL dataset with one of its neighborings randomly, and rendered them using two importance sampling techniques(ours, cosine).

ImportanceSampling-interpolation-envmap

Hypernet Interpolation

To validate the interpolation ability of our hypernet, we randomly interpolated the materials, similar to the handling in importance sampling. Then we compared the renderings of the BRDFs reconstructed using our NPs model and our hypernet.

HypernetInterpolation